Any AI model is largely as reliable as the accuracy of the patterns embedded in its training dataset. When it comes specifically to medical AI, the need for its reliability skyrockets to the maximum, because behind every evaluation of an X-ray image stands a real diagnostic decision. To avoid making mistakes under such conditions, the model must rely on a clear benchmark – examples that demonstrate the single correct interpretation of what is seen.
In machine learning, such benchmark answers are called ground truth. But before using them as a yardstick of correctness, it is necessary to understand where they derive their credibility. After all, behind conversations about "the reliability of a model trained on ground truth" lies a fundamental question: who exactly defines what constitutes clinical truth?
Let us return to the X-ray image at the moment when it is being evaluated by doctors rather than the model. One radiologist sees a small area that might indicate a pathology. Another sees the exact same signs but considers them benign. Both have professional training, yet their interpretations differ.
In real-world medicine, such uncertainty is a reason for an additional examination. For the team preparing the data for medical AI, it creates a specific problem: which answer should be recorded in the dataset?
In a study published in the British Journal of Radiology, six radiologists independently evaluated 1100 chest X-raysacross six features. The consistency of their conclusions depended heavily on the specific feature they were looking for.
The researchers also tested whether discussing disputed cases would help improve the ground truth labels. Additional review increased consistency for certain features, but did not eliminate all discrepancies.
Image credit: https://pmc.ncbi.nlm.nih.gov/articles/PMC8248225/
Naturally, in machine learning, one would always like to receive a single correct answer that the model can rely on. However, if even experts frequently interpret medical data differently, forcibly reducing all cases to a single label can conceal important nuances of the clinical picture. Ultimately, the credibility of the benchmark training dataset depends on how well-reasoned the team's strategy is for resolving disputed cases.
One way to establish ground truth is to rely on clinical data that helps confirm or refute the initial assumption. This could include laboratory test results, a histological conclusion, a confirmed diagnosis, or information about the patient's subsequent disease progression.
However, not every patient undergoes additional examinations, results from different tests may contradict one another, and the final diagnosis does not always directly correspond to what the model is supposed to learn to recognize. If a model is supposed to detect a specific feature in an image, a confirmed diagnosis may be useful, but obviously insufficient, for instance, to determine its exact boundaries.
This approach helps reduce reliance on the subjective assessment of a single specialist, yet it does not eliminate all problems. Even when a medical history is available, the team must determine which specific information confirms the required feature, which merely complements the picture, and which should not influence a specific label.
This is particularly noticeable in projects where a single patient's data spans multiple related examinations. For example, while working with ultrasound data, our team had to maintain annotation consistency across the entire medical history. To be frank, this is an exceptionally complex type of data interaction. If an error was detected in even a single sample, the team returned to other related data for that patient to check whether contradictions had arisen elsewhere in the labeling.
Clinical evidence can undeniably serve as a solid foundation for the ground truth, but only when the team clearly understands exactly what it confirms and whether that aligns with the model's future application.
An indispensable and reliable method, of course, is to involve multiple specialists, compare their conclusions, and determine which answer best fits the available evidence and task criteria.
The simplest option is to select the answer supported by the majority. If two out of three experts label a specific feature as present, the dataset can receive the corresponding label. This helps standardize decisions, but it carries an obvious limitation: the majority can also be wrong.
Therefore, in more complex cases, a targeted review of disputed cases is applied. Experts compare their arguments, consult additional clinical data, or bring in additional specialists with relevant expertise. The goal of such a process is to clarify why the discrepancy arose and whether it can be resolved according to agreed-upon criteria.
For example, during the annotation of medical images of the diaphragm on fluoroscopy and X-rays, we encountered a situation where on some images its contours were traced quite clearly, while on others the boundary between the structure and surrounding tissues was so faint that it remained ambiguous even after several rounds of discussions.
Description: An example of medical image annotation, where subtle anatomical boundaries require consistent labeling rules and expert review.
Here, our team worked in small batches: medical specialists on the client side reviewed the results, and their feedback was used to refine guidelines and calibrate labeling approaches.
As a result, consensus helps establish a shared assessment where possible. But the pursuit of a single definitive answer has its limits: in medicine, expert agreement does not always mean that doubts no longer exist.
However, not every disagreement among experts needs to be eliminated by picking a single definitive answer. Sometimes it is more useful to preserve the fact of that disagreement itself. For this purpose, probabilistic labels are used to reflect the degree of unanimity among the specialists' evaluations.
For example, four experts are asked to evaluate a single tissue sample. Three believe it shows signs of atypia, while the fourth sees insufficient grounds for such a conclusion. Instead of leaving only the answer "atypia" in the dataset or conducting an additional round of discussion, the team can record the distribution of evaluations: 0.75 for "atypia" and 0.25 for "no signs of atypia".
During training, the model takes both evaluations into account: the primary benchmark for it becomes the presence of atypia, but the disagreement between experts also remains part of the training signal. In medical annotation, it is especially important to distinguish between what is explicitly stated in the dataset and what can only be inferred. This approach is particularly useful where the boundary between categories is blurred, and excessive model confidence is undesirable.
The main value of probabilistic labels is precisely that they do not force the dataset to pretend that the model can rely on it with certainty when making a conclusion. However, in real-world application, this depends on whether the model can reliably assess its own uncertainty and whether its predictions have been verified in clinically relevant conditions.
From our experience in data preparation for AI, all three approaches can be useful, but none of them guarantees a reliable ground truth on its own. In fact, its quality is determined at every step of the data preparation process.
It begins with defining what exactly the model is being created for. Is it for detecting specific features in an image, extracting information from clinical notes, classifying medical cases, or predicting treatment outcomes? What data can be considered representative, which labeling rules must be established, and what expertise should be brought in depends on the task set before the model.
Sometimes difficulties arise precisely because the definition of the correct label changes along with the understanding of the task. In one dental project where we worked with images from clinical and mock-up conditions, labeling requirements were refined throughout the work based on changes in probe logic. Our team had to update the classification of molars and premolars taking quadrants into account, clarify permissible boundaries for contour determination, and add new attributes – for example, signs of probe contact with tissues. Here, it is crucial to monitor whether previous labels remain compatible with the new logic and to avoid situations where the same anatomical structure is labeled differently at different stages of the project.
The composition of the data itself also plays a critical role. Even flawlessly annotated examples cannot compensate for the absence of key data categories. Benchmark ground truth demands diversity: data should come from scanners made by different manufacturers, cover patients across different ages and ethnicities, and contain a sufficient number of complex, rare cases.
It is also worth noting that annotating complex DICOM or WSI images, which frequently accompany medical projects, requires specialized tools. For this, teams usually use platforms developed with medical annotation needs in mind, particularly Encord, RedBrick AI, and V7 Darwin. At Keymakr, we more frequently work with V7 and our proprietary Keylabs platform, but we usually explicitly coordinate the choice of the labeling environment, as this significantly impacts the speed of the annotation process, its quality, and cost. Medical projects often have strict deadlines, so this decision needs to be made with the model's real needs in mind.
The next step is detailed annotation guidelines. If a rule can be interpreted in multiple ways, the team must catch this before discrepancies spread across thousands of examples. To prevent this, for instance, we apply multiple levels of control in our project workflow: verification of completed tasks, internal quality checks, and, as required by the project, client validation.
Ultimately, as a company that prepares data for training models, we are used to seeking and shaping an absolute benchmark in everything. But the truth is that in medicine, the most reliable ground truth is one that evolves alongside new clinical evidence, updated protocols, and the models for which it is created. In this sense, it can be compared to a part of a living data pipeline that requires just as much attention and control as the model itself.